NBER WORKING PAPER SERIES

THE LONG-TERM CONSEQUENCES OF VIETNAM-ERA CONSCRIPTION ANDGENOTYPE ON SMOKING BEHAVIOR AND HEALTH

Lauren SchmitzDalton Conley

Working Paper 21348http://www.nber.org/papers/w21348

NATIONAL BUREAU OF ECONOMIC RESEARCH1050 Massachusetts Avenue

Cambridge, MA 02138July 2015

Dalton Conley would like to thank the Russell Sage Foundation for financial support for this project.The Health and Retirement Study (HRS; accession number 0925-0670) is sponsored by the NationalInstitute on Aging (grant numbers NIA U01AG009740, RC2AG036495, and RC4AG039029) andis conducted by the University of Michigan. Additional funding support for genotyping and analysiswere provided by NIH/NICHD R01 HD060726. The views expressed herein are those of the authorsand do not necessarily reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies officialNBER publications.

© 2015 by Lauren Schmitz and Dalton Conley. All rights reserved. Short sections of text, not to exceedtwo paragraphs, may be quoted without explicit permission provided that full credit, including © notice,is given to the source.

The Long-Term Consequences of Vietnam-Era Conscription and Genotype on Smoking Behaviorand HealthLauren Schmitz and Dalton ConleyNBER Working Paper No. 21348July 2015JEL No. I1,I12,I14

ABSTRACT

Research is needed to understand the extent to which environmental factors mediate links betweengenetic risk and the development of smoking behaviors. The Vietnam-era draft lottery offers a uniqueopportunity to investigate whether genetic susceptibility to smoking is influenced by risky environmentsin young adulthood. Access to free or reduced-price cigarettes coupled with the stress of military lifemeant conscripts were exposed to a large, exogenous shock to smoking behavior at a young age.Using data from the Health and Retirement Study (HRS), we interact a genetic risk score for smokinginitiation with instrumented veteran status in an instrumental variables (IV) framework to test forgenetic moderation (i.e. heterogeneous treatment effects) of veteran status on smoking behavior andsmoking-related morbidities. We find evidence that veterans with a high genetic predisposition forsmoking were more likely to become regular smokers, smoke heavily, and are at a higher risk ofbeing diagnosed with cancer or hypertension at older ages. Smoking behavior was significantlyattenuated for high-risk veterans who attended college after the war, indicating post-service schoolinggains from veterans’ use of the GI Bill may have reduced tobacco consumption in adulthood.

Lauren SchmitzNew School for Social ResearchNew York, NY 10011schml347@newschool.edu

Dalton ConleyNew York University249 West 29th Street #2ENew York, NY 10001-5230and NBERdaltonclarkconley@gmail.com

2

1. Introduction

Cigarette smoking continues to be a significant and costly public health problem in the

United States. Cigarette smoking and exposure to tobacco smoke resulted in at least 443,000

premature deaths, approximately 5.1 million years of potential life lost, and $96.8 billion in

productivity losses annually between 2000-2004 (CDC, 2008). Studies on the persistence of

smoking across the life span have consistently found heavy smoking in young adulthood to be a

strong predictor of smoking later in life (Merline, O'Malley, Schulenberg, Bachman, & Johnston,

2004). Moreover, few people begin smoking after adolescence (Kandel & Logan, 1984), and

young adults who manage to quit smoking are generally able to avoid any significant lifetime

health consequences (U.S. Department of Health & Human Services, 1990). These findings

have motivated health policymakers and researchers to understand the complex influences that

contribute to smoking initiation and cessation during this critical time period. However, the vast

majority of research has telescoped its efforts around understanding either the social or the

biological causes of early onset cigarette use, and has therefore neglected the extent to which

environmental factors may moderate or mediate links between genetic risk and the development

of smoking behaviors.

The Vietnam-era draft lottery offers a unique opportunity to investigate whether genetic

predisposition to smoking is influenced by risky or stressful environments in young adulthood.

Between August 1964 and May 1975 approximately 2.7 million Americans were inducted in the

military—1,857,304 of which entered military service through the Selective Service System

(U.S. Department of Veteran Affairs, 2015; U.S. Selective Service, 2015). Whether they were

deployed to Vietnam or served outside the theater of war, conscripts were exposed to a large,

exogenous shock to smoking behavior when they were approximately 19-22 years of age. All

3

men who served had access to tax-free cigarettes at military bases and commissaries, and men in

combat received free cigarettes as part of their K- and C-rations (U.S. Department of Health and

Human Services, 1989, pp. 425-426). Given the high concentration of extant smokers at military

bases and facilities, peer effects may have further augmented the effect of cheap or reduced-price

cigarettes (Eisenberg & Rowe, 2009). In addition, many servicemen were subjected to stressful,

if not traumatic experiences, which could have led to elevated levels of depression and anxiety—

both of which are associated with an increased risk of smoking (Glassman et al., 1990;

Winefield, Winefield, Tiggemann, & Goldney, 1989). While a few studies have used the draft

lotteries to evaluate the causal impact of military service on smoking behavior and health, this

study is the first to examine whether genetic susceptibility to smoking is influenced by military

service in young adulthood.

To properly test for gene-by-environment (GxE) interactions between cigarette

consumption and military service, we incorporate a polygenic score or genetic risk score (GRS)

for smoking behavior into an instrumental variables (IV) framework. Specifically, because

selection into the military is far from random, and likely to be correlated with factors like

socioeconomic background or underlying genotype, we follow prior studies and use respondent

date of birth files for cohorts born between 1948 and 1952 in the Health and Retirement Study

(HRS) as an instrumental variable to code for draft eligibility. Instrumented veteran status is

then interacted with a GRS for smoking initiation to test for genetic moderation, or

heterogeneous treatment effects, of veteran status on smoking behavior and smoking-related

morbidities. Existing efforts to find associations between genetic variation and social behavior

in large, multidisciplinary surveys are often unable to support causal inference because they use

endogenous measures of the environment, genotype, or both (Conley, 2009; Fletcher & Conley,

4

2013). By exploiting natural variation in exposure to military service, this study lends insight

into how physiological pathways might be moderated or mediated by environmental influences.

Our results show veterans with a high genetic predisposition for smoking—defined as

having a GRS one or two standard deviations above the mean—were more likely to become

smokers, or smoke more than 100 cigarettes in their lifetime, and smoke in greater quantities on

a day-to-day basis than comparable non-veterans with the same GRS. Veterans with a GRS one

to two standard deviations above the mean were approximately 57 to 71 percent more likely to

become smokers, and smoked 18 to 27 cigarettes more per day than comparable non-veterans.

Critically, we find these effects were significantly attenuated for high-risk veterans who attended

college after the war, indicating post-service schooling gains from veterans’ use of the GI Bill

may have reduced tobacco consumption in adulthood. We also find some evidence that high-risk

veterans were more likely to be diagnosed with cancer and have higher mean arterial blood

pressure (MAP) than comparable non-veterans.

The study is structured as follows. After reviewing the relevant literature on the genetics

of smoking behavior and its relation to the draft lottery and military service in Section 2, we

present an in-depth explanation of the data and methodology behind our causal GxE interaction

model in Sections 3 and 4. Finally, Section 5 gives a detailed explanation of our results, and

Section 6 concludes.

5

2. Background

2.1 The Genetics of Smoking Behavior

Twin studies on the genetics of smoking behavior have found genetic influences play an

important role in smoking initiation, cessation, and response to antismoking interventions.

Studies estimate heritability accounts for approximately 44 percent of the variation in the

frequency of smoking (Slomkowski, Rende, Novak, Lloyd-Richardson, & Niaura, 2005) and 65

percent of the variation in current and lifetime tobacco use (Maes et al., 1999). Recently,

genome-wide association studies (GWASs) of current and former smokers have identified

genome-wide significant loci that are associated with heavy smoking in adulthood (Furberg et

al., 2010; Liu et al., 2010; Thorgeirsson et al., 2010).

Importantly, factors that influence smoking initiation appear to differ from factors that

contribute to the persistence and quantity of smoking behavior. In particular, shared

environmental influences play a larger role in smoking initiation than persistence, and there is

considerable evidence that peer effects fuel onset in adolescence. However, only a few studies

have integrated environmental and biological explanations for cigarette use (Boardman, Saint

Onge, Haberstick, Timberlake, & Hewitt, 2008; Daw et al., 2013; Kendler, Gardner, Jacobson,

Neale, & Prescott, 2005; Timberlake et al., 2006). For example, using twin data from the

National Longitudinal Study of Adolescent Health, Boardman et al. (2008) find the heritability

of daily smoking is higher in schools where the most popular students are smokers and reduced

in schools where the majority of students are non-Hispanic and white. On the other hand, using

findings from GWASs to derive a GRS for smoking quantity, Belsky et al. (2013) find high

genetic risk leads to persistent heavy smoking, nicotine dependence, and difficulty quitting for

6

individuals who progress rapidly from smoking initiation to heavy smoking during adolescence.

The authors conclude interventions that disrupt the socially developmental progression of

smoking behavior in young adulthood may mitigate the development of adult smoking problems

for individuals with high genetic risk.

These findings signal the need for further research that can elucidate specific

environmental mechanisms that moderate or mediate links between genetic risk and the

development and persistence of smoking behaviors. This study is the first to use a polygenic

score to measure genetic risk and a plausibly exogenous environmental exposure—the Vietnam-

era draft lotteries—to identify whether military service shaped tobacco consumption across the

life span.

2.2 The Draft Lottery as a Quasi-Natural Experiment

Between December 1969 and February 1972, the U.S. Selective Service held four

Vietnam-era draft lotteries. Each of these draft lotteries randomly assigned men in eligible birth

cohorts ordered induction numbers through a hand drawing of birthdates. Individual birthdates

(including February 29th) were assigned a corresponding number between 1 and 366, placed in a

blue capsule, and then individually drawn until each day of the year was paired with a random

sequencing number (RSN). Depending on the U.S. Department of Defense’s needs for

manpower, a cut-off number for draft eligibility was chosen for the following calendar year.

Men with RSNs below the cut-off were considered draft-eligible for that particular year and were

therefore more likely to serve, whereas men with RSNs above the cutoff were not at risk of

conscription. The first lottery was held on December 1, 1969 for men born between 1944 and

1950; subsequent lotteries drafted men from the 1951 and 1952 birth cohorts, respectively. A

7

final lottery was held in 1972 for men born in 1953, but since no new conscription orders were

issued after 1972, these men were never called to service.

The random assignment mechanism of the draft lotteries has been used to identify the

effects of war-time military service on a host of outcomes, including economic (Angrist, 1990;

Angrist & Chen, 2011; Angrist, Chen, & Song, 2011), family (Conley & Heerwig, 2011;

Heerwig & Conley, 2013) and health outcomes (Angrist, Chen, & Frandsen, 2010; Conley &

Heerwig, 2012; Dobkin & Shabani, 2009). While earlier studies uncovered small but statistically

significant effects of military service on mortality (Hearst, Newman, & Hulley, 1986) and

earnings (Angrist, 1990), by and large studies have been unable to uncover a causal link between

conscription and long-term health outcomes, lifetime earnings losses, employment, or labor force

participation.

To our knowledge, only a few studies have estimated the causal effects of military

service on smoking behavior and health. Bedard and Deschênes (2006) estimate the long run

effects of military service in World War II and the Korean War on smoking, mortality, and other

measures of health. Since detailed individual-level data is not available for these veterans over

time, they construct panels of all-cause and cause-specific mortality rates for different birth

cohorts and use instrumental variable approaches that exploit variation in the proportion of men

who served across birth cohorts. They find military service during WWII and the Korean War

era is associated with a 30 percentage point increase in the probability of lifetime smoking, and

36 to 79 percent of veteran deaths from heart disease and lung cancer are attributable to military-

induced smoking.

Using the Vietnam-era draft lotteries, Dobkin and Shabani (2009) find no evidence of

increased smoking behavior in the National Health Interview Survey (NHIS) over the years

8

1997-2004, indicating any increase in smoking behavior due to conscription in young adulthood

likely dissipated by mid-life, or by the time veterans were aged 45-54. They also do not find

evidence that Vietnam-era service resulted in poorer health outcomes as these populations age, or

across three successive time periods in the NHIS (1974-1981, 1982-1996, and 1997-2004).

Similarly, Eisenberg and Rowe (2009) find Vietnam-era military service increased the

probability of smoking by 35 percentage points when veterans were aged 25 to 30, or as of 1978-

1980 in the NHIS, but find any early effects were substantially attenuated later in life, and did

not result in any long-term health consequences.

A major limitation of these studies is they are only able to estimate the average treatment

effect of conscription on smoking behavior and health. In other words, studies have not been

able to account for heterogeneous treatment effects by genotype, or the possibility that military-

induced smoking varies across the spectrum of genetic risk. In addition, a significant link

between military service and smoking behavior may be difficult to unearth due to the potentially

offsetting effect of the GI Bill, which encouraged veterans to attend college after service. Prior

research has shown the GI Bill increased schooling for WWII (Bound & Turner, 2002), Korea-

era veterans (Stanley, 2003), and Vietnam veterans (Angrist & Chen, 2011). Among white

Vietnam veterans, 50 percent used funds from the GI Bill for education—63 percent of which

was put towards college courses, resulting in 0.26 more years of college on average compared to

non-veterans (Angrist & Chen, 2011). Higher educational attainment is strongly associated with

better health (e.g. Fuchs, 1982; Grossman, 2006; Lleras-Muney, 2005), and may have exerted an

important influence on veterans by reinforcing positive health behaviors—either through

improved decision-making and access to health-related information or because the boost in

future earnings potential incentivized better health investments (Grossman, 1972). Along these

9

lines, numerous studies have found a causal link between education and a reduction in smoking

behavior (e.g. de Walque, 2010; Farrell & Fuchs, 1982). In particular, studies have used the

draft avoidance behavior documented by Card and Lemieux (2001) to infer causation from

education to smoking. For example, De Walque (2007) and Grimard and Parent (2007) exploit

the uptick in male college attendance in the early 1960s that occurred as a result of educational

deferments and find convincing evidence that educational attainment reduced tobacco

consumption.

Given the strong negative correlation between education and smoking in the literature,

the GI bill may have acted as a second moderating environmental lever after the war, particularly

for high-risk individuals. On the other hand, educational attainment may be endogenous in the

GxE smoking model if it is driven by factors outside of the draft, such as socioeconomic

background or unmeasured ability, that influence health behaviors. However, if draft status was

orthogonal to standard socio-demographic variables at the time of the lottery and educational

attainment after the war was directly related to draft eligibility, as past studies have shown,

education may be another mechanism affecting the smoking behavior of conscripts. As a result,

we also explore whether obtaining a college or advanced degree after the war moderated the

smoking behavior of high-risk individuals in separate GxE interaction models.

3. Data and Descriptive Statistics

Our primary data source is the Health and Retirement Study (HRS). The HRS is a

nationally representative, longitudinal dataset of individuals aged 51 plus and their spouses that

was launched in 1992. It follows individuals from the time of their entry into the survey until

their death. It introduces a new cohort of participants every six years, and interviews around

10

20,000 participants every two years. For the purposes of this study, the HRS genotyped 12,507

respondents in 2009 that provided DNA samples in 2006 and 2008. Genotype data is collected

alongside detailed health, economic, and social data, including information about chronic illness,

functional ability, depression, self-assessed health status, health related behaviors, and veteran

status.

The majority of men who were draft eligible during the Vietnam era entered the HRS in

2004 as part of the “Early Baby Boomer” cohort—a nationally representative sample of men and

women born between 1948-1953. However, to maximize sample size, we also include spouses

of female respondents from former cohorts that were born between 1948-1952. After quality

control analysis on the genotype data, 12,150 observations remained (see section 4.1 for QC

specifics). We then exclude respondents who report being female (N=7,176), black (N=574),

American Indian, Alaskan Native, Asian, or Pacific Islander (N=103), or who were born before

1948 (N=3,450) or after 1952 (N=216). Our final sample consists of 631 white men born

between 1948-1952 who provided DNA samples in 2006 or 2008.

3.1 Smoking Phenotypes

The HRS asks several questions that pertain to current and former smoking behavior. All

respondents were asked if they ever smoked more than 100 cigarettes in their lifetime (not

including pipes or cigars). If the respondent answers yes to this question, they are asked if they

currently smoke and if so, how many cigarettes or packs of cigarettes they smoke per day. If

they report smoking in their lifetime but do not currently smoke, they are asked how many

cigarettes or packs they smoked per day when they were smoking the most. The HRS also asks

all current and former smokers when they started smoking, and if they no longer smoke, when

11

they stopped. Smoking questions were fielded to both current and former smokers in the 1992

and 1998-2010 waves. 1

Our primary phenotypes of interest for smoking behavior include a

dichotomous variable for whether or not an individual reports smoking 100 cigarettes or more in

their lifetime (“Ever Smoker”) and a continuous variable for either the amount of cigarettes the

respondent reports smoking per day if they currently smoke or the maximum amount of

cigarettes the respondent smoked per day if they no longer smoke (“Cigarettes Per Day”).

There are a few limitations with the measurement of both phenotypes that bear some

acknowledgement. First, with respect to the “Ever Smoker” phenotype, the majority of current

or former smokers in our sample (74.4%) report smoking before age 19, or before they were

draft eligible—i.e. most men in our sample who served in the military did not initiate smoking

during their tour of service or when they were draft eligible (only 16.1% report onset between

the ages of 19 and 22). However, because we do not have any information on the frequency of

smoking behavior at younger ages, it is impossible to discern at what point individuals crossed

the 100-cigarette threshold. Tobacco use changes considerably during adolescence and young

adults smoke fewer cigarettes daily and are less likely to smoke regularly than adults (CDC

2003; Johnston, O'Malley, Bachman, & Schulenberg, 2005; Rouse, Sanderson, & Feldmann,

2002). Thus, nondaily, experimental, or “social” smoking behavior early on may have evolved

into established use as a result of military service. Likewise, men who started smoking after age

22 (9.4%) may have initiated because they experienced high levels of stress, difficulties

assimilating back into civilian life, or other negative downstream effects from military service

after the war. In both cases, because we cannot rule out the possibility that Vietnam-era service

1 In 1994 and 1996, the smoking questions were fielded to current smokers only. However, since

questions about former smoking behavior were asked in 1992, and a new cohort was not added until

1998, data on past smoking behavior is available for the majority of participants.

12

was a gateway to longer-term or more persistent tobacco use, we include the “Ever Smoker”

phenotype in our results with the caveat that some men may have smoked 100 cigarettes or more

before the war.

Second, since the HRS only asks former smokers about their lifetime maximum CPD, our

“Cigarettes Per Day” phenotype measures current CPD for men who still smoke (approximately

45.2% of current and former smokers in our sample), and maximum CPD for former smokers.

Current daily smoking may be an inaccurate measure for lifetime intensity. For example, it is

unlikely that a person who currently smokes two packs a day has been consistently smoking this

amount since they started. Smokers may have changed their lifetime smoking behavior for

health reasons, to compensate for the changing nicotine yields in manufactured cigarettes, or to

adjust for increases in the real price of cigarettes over time (CDC, 1994). The decreasing social

acceptability of smoking may also cause current smokers—and heavy smokers in particular—to

underreport the number of cigarettes they smoke.2 A comparison of sample means finds there is

no significant difference between self-reports of CPD for current and former smokers—i.e.

current smokers are not smoking more on average than former smokers did at their peak. Thus,

while the ideal measure would report maximum CPD for all respondents, if smoking behavior

among current smokers tends to remain at or below peak levels, we would expect our results to

be biased downward if anything as a result of this discrepancy.

2 In general, self-reports of smoking behavior have been shown to be a reliable measure over time despite

mounting social stigma (Hatziandreu, Pierce, Fiore, Grise, Novotny, & Davis, 1989) and findings from

biochemical validation studies suggest self-reported usage is a valid estimate of smoking status in the

population (Fortmann, Rogers, Vranizan, Haskell, Solomon, & Farquhar, 1984; Patrick, Cheadle,

Thompson, Diehr, Koepsell, & Kinne, 1994).

13

3.2 Smoking-Related Morbidities

Genetically-at-risk smokers tend to not only smoke more, but may also experience

increased toxin exposure at fixed quantities of smoking, and are at increased risk for heart

disease and lung cancer (Le Marchand et al, 2008; Thorgeirsson et al., 2008; Spitz, Amos,

Dong, Lin, & Wu, 2008). To see if smoking among higher-risk genotypes resulted in any long-

term health consequences, we examine outcomes for two primary smoking-related health

phenotypes: cancer and hypertension.3 We use a dichotomous variable for incidence of cancer,

which is equal to “1” if the respondent reports ever being diagnosed with cancer or a malignant

tumor of any kind except skin cancer and “0” otherwise. For hypertension, we calculate mean

arterial blood pressure (MAP), or the average arterial pressure during a single cardiac cycle.

MAP is often used as an alternate indicator of blood flow and is believed to be a better indicator

of tissue perfusion than systolic blood pressure (SBP) since it accounts for the fact that two thirds

of the cardiac cycle are spent in diastole. 4

MAP has also been shown to be a significant

independent predictor of all-cause, cardiovascular, and coronary mortality (e.g. Benetos, Safar,

Rudnichi, Smulyan, Richard, Ducimetière, & Guize, 1997). To calculate MAP in the HRS

sample, we use measures of systolic and diastolic blood pressure taken during enhanced face-to-

face interviews with an automated sphygomanometry (inflated blood pressure cuff).

3 We also explored outcomes related to lung disease and lung function but failed to find any significant

effects. Results are available from the authors upon request.

4

14

3.3 Draft Eligibility

To assess whether an individual was draft eligible, we follow prior studies that use the

Vietnam-era draft lotteries as a quasi-natural experiment, and assign order of induction numbers

based on restricted respondent date of birth files for cohorts born between 1948 and 1952 in the

HRS. Table 1 shows the draft eligibility cutoffs and birth cohorts affected by each draft lottery.

Based on the eligibility cutoffs reported, draft eligibility is equal to “1” if the lottery number

corresponding to a respondent’s date of birth was called in the corresponding draft year and “0”

otherwise.5

Table 1. Draft eligibility numbers by birth cohort and lottery year

Lottery Year Birth Cohort(s) Eligibility Ceiling

1969 1944-1950 195

1970 1951 125

1971 1952 95

1972 1953 --

Source: U.S. Selective Service.

Due to the high number of men who volunteered for the military or received educational

deferments before the first draft lottery, the vast majority of studies exclude the older 1944-1949

birth cohorts, since this sample is far from representative (Angrist, 1990). For example, using

the NHIS, Dobkin and Shabani (2009) regress veteran status on draft eligibility and find men

born between 1950-1952 on draft eligible days were 14.8 percentage points more likely to serve,

whereas the probability of service for draft eligible men from the 1948 or 1949 cohorts was only

around 7 percent. We replicate this analysis with the HRS sample and do not find a significant

5 The results of the Vietnam draft lottery are available at: http://www.sss.gov/lotter1.htm.

15

difference in the likelihood of service for draft eligible men in the 1948-1952 and 1950-1952

birth cohorts.6 Thus, because our sample of genotyped veterans is significantly lower than past

studies that use larger, nationally representative data sets like the NHIS, we include the 1948-

1949 birth cohorts to maximize the power of this study.

3.4 Descriptive Statistics

Table 2 reports descriptive statistics for the sample by veteran and draft eligibility status.

Independent t-tests comparing the means of variables between groups reveal veterans differ

substantially from non-veterans along several dimensions. Veterans are more likely to become

regular smokers (0.692) than non-veterans (0.577), smoke 4.327 more cigarettes per day on

average, and are 9.6 percent more likely to continue smoking at older ages. Veterans have

higher rates of high school completion (0.606) than non-veterans (0.418), but were less likely to

obtain a college or advanced degree; only 23.2 percent of veterans completed college compared

with 40.4 percent of non-veterans. On the other hand, descriptive statistics show no significant

difference in marital status or smoking-related morbidities between veterans and non-veterans.

Importantly, since the HRS is a study of older adults and we are observing veterans

nearly 40 years after the Vietnam war, the exogeneity of the GRS and the random nature of the

draft eligible and ineligible populations may be called into question if there is substantial attrition

due to poor health or selective mortality in the Vietnam cohort. For example, potential problems

could arise if conscripts who survived to be genotyped in 2006 or 2008 have lower GRSs or are

more likely to be educated and therefore have lower mortality rates than comparable populations.

6 Specifically, we regress veteran status on draft eligibility and a constant with controls for month of birth.

Men born between 1948 and 1952 were 15.7 percentage points more likely to serve, while men born

between 1950 and 1952 were 15.3 percentage points more likely to serve. Results are available from the

authors upon request.

16

In both cases, if selective mortality is correlated with the GRS or with educational attainment,

draft eligibility is no longer exogenous, and the exclusion restriction on which our 2SLS

estimates rest will be violated.

To explore this further, we compare descriptive statistics for the draft eligible and

ineligible populations. The results in Table 2 reveal no significant differences at the mean

between draft eligible and ineligible men in the HRS by GRS, smoking status, health,

demographics, or education. This suggests selection bias is modest if nonexistent in the aging

Vietnam veteran sample—a finding that is supported in the literature. Conley and Heerwig

(2012) find no effect of draft exposure on mortality, including for cause-specific death rates, in a

larger sample of national death records, and Angrist et al. (2010) find little evidence of elevated

mortality among draft-eligible men in the 2000 census. Importantly, Conley and Heerwig (2012)

do find some evidence of elevated mortality among draft-eligible, college-educated men.

However, they argue this effect works in the opposite direction of putative education-enhancing

effects that could potentially violate the exclusion restriction in IV regression models—i.e. high

SES men were not more likely to survive.

Finally, Table 3 compares descriptive statistics for white men born between 1948-1952

who provided genetic data (N=631) with same-age peers who were not genotyped (N=540). By

and large there are few differences between the two groups. Genotyped men report smoking 2.4

more cigarettes per day, are 5.1 percent less likely to be Hispanic, and are 4.1 percent more

likely to report being married or partnered than non-genotyped men.

17

4. Methodology

To properly test and identify GxE interaction effects, we incorporate the latest

approaches from population genetics into an IV framework to ensure both G and E are

independent of each other. Existing efforts to find associations between genetic variation and

social behavior in large, multidisciplinary surveys are often unable to support causal inferences

because they use endogenous measures of the environment, genotype, or both (Conley, 2009;

Fletcher & Conley, 2013). In the case of military service, the descriptive statistics in Table 2

reveal selection into the military is far from random, and likely to be correlated with factors like

socioeconomic background, prior health status, or genotype. Given these differences, it would

be impossible to sort out the effects of military service from the effects of other GxE or gene-

gene (GxG) interactions in a model that uses self-reported veteran status to estimate GxE effects.

In addition, because smoking behavior is highly polygenic, or reflects small contributions

from many genetic loci, the use of single genetic variants in our GxE model may result in a form

of omitted variable bias, whereby crucial genetic information is obscured and left sitting in the

error term, confounding estimates. To yield a comprehensive measure of genetic risk, we

calculate a genetic risk score (GRS) that sums risk alleles across independently segregating

genotyped single nucleotide polymorphisms (SNPs). This condenses the amount of genetic

information available for an individual into a single, quantitative measure of genetic risk, thus

minimizing the possibility that “G” is acting as a proxy for other GxG, ExE, or GxE interactions

while also maximizing statistical power. This approach also has the added advantage of using

main effect analysis already extant in the literature, which benefits from large consortia of

adequately powered data to efficiently estimate individual allelic effects.

18

4.1 Genetic Risk Score

We calculate linear polygenic scores for the HRS sample based on results from a recent

GWAS meta-analysis conducted across 16 studies by the Tobacco and Genetics (TAG)

Consortium among people of European ancestry (Furberg et al., 2010). Specifically, we

construct a composite “risk score” or genetic risk score (GRS) composed of weighted effects of

specific single nucleotide polymorphisms (SNPs) for both smoking initiation (ever versus current

smoker) and smoking quantity (number of cigarettes per day or CPD) using the HRS Genotype

Data Version 1 (2006-2008 samples).7 We analyzed original genotype data to construct the

score—i.e. imputed data was not analyzed. SNPs in the HRS genetic database were matched to

SNPs with reported results in the GWAS. In the HRS genetic data set, 802,191 SNPs were

available to construct the smoking initiation score and 802,903 SNPs were available to construct

the CPD score. The matched set of SNPs was “pruned” to account for linkage disequilibrium

(LD) using the clumping procedure (which considers the level of association between the SNP

and the phenotype, not simply LD) in the second-generation PLINK software (Chang, Chow,

Tellier, Vattikuti, Purcell, & Lee, 2015; Purcell & Chang, 2014).8 For each of these SNPs, a

loading was calculated as the number of smoking-associated alleles multiplied by the effect-size

estimated in the original GWAS. We do not impose a p-value threshold; instead, SNPs with

relatively large p-values or small effects are down weighted in the composite score. Loadings

are then summed across the SNP set to calculate the polygenic score. After LD

7 Genotyping was performed on the HRS sample using the Illumina Human Omni-2.5 Quad beadchip

(HumanOmni2.5-4v1 array). The median call-rate for the 2006-2008 samples is 99.7%.

8 Clumping takes place in two steps. The first pass is done in fairly narrow windows (250kb) for all SNPs

(the p-value significance thresholds for both index and secondary SNPs is set to 1) with a liberal LD

threshold (r2=0.5). In a second pass, SNPs remaining after the first prune are again pruned in broader

windows (5000kb) but with a more conservative LD threshold (r2=0.2).

19

pruning/clumping was applied, 178,834 SNPs were used to generate the smoking initiation score

and 179,245 SNPs were used to generate the CPD score.

The GRSs are standardized to have a mean of 0 and a standard deviation of 1 for the

population of white males born between 1948 and 1952. Both the smoking initiation and

smoking quantity scores are predictive of our two smoking phenotypes (cigarettes per day and

ever smoker) in base, non-interactive main effects models for the entire sample of white,

genotyped males in the HRS (see Table 4). However, in our sample of white males born

between 1948 and 1952, the smoking quantity or CPD score is insignificant. Thus, we only

report results from the GxE interaction analysis that use the GRS for smoking initiation.9

To control for confounding by population stratification—or the non-random distribution

of genes across populations—we use principal components. The principal components measure

the uncorrelated variation or dimensions in the data, accounting for ethnic or racial differences in

genetic structures within populations that could bias estimates. We calculate the principal

components from the entire sample of genotyped respondents in the HRS, and include the first

20 in our regression analysis—a dimensionality that has generally proven adequate in the

literature (Price et al., 2006). Controlling for the first 20 principal components accounts for any

systematic differences in ancestry that can cause spurious correlations while also maximizing the

power that is needed to detect true associations.

To test the sensitivity of the genetic score to the environmental effect we regress smoking

status on the smoking initiation GRS in the draft eligible and ineligible populations, and find

preliminary evidence of a GxE effect (see Table 5); the smoking initiation score is positive and

predictive of all four phenotypes in the draft eligible population, whereas the impact of the GRS

is insignificant and considerably reduced in magnitude in the draft ineligible population. Yet,

9 GxE interaction models with the CPD GRS are available from the authors upon request.

20

commensurate with previous findings, we are unable to detect a main effect of our environmental

exposure, or instrumented veteran status, on smoking behavior and health in the base 2SLS

model (see Table 6). The significant gradient between the genotype score and smoking behavior

in the draft eligible population may explain the absence of an average treatment effect—i.e. other

studies may have been unable to detect a main effect precisely because outcomes vary

considerably by genotype within the draft eligible population.

4.2 Empirical Model

To account for selectivity issues due to endogenous sorting into the military or

unobserved heterogeneity that may bias results, we use an instrumental variables (IV) approach

to properly identify GxE effects. Following prior studies, we use the Vietnam-era draft lottery as

an instrument for veteran status to identify the causal effect of military service and genotype on

veterans’ smoking behavior and health after age 50. Since past research has shown draft

eligibility is 1) highly correlated with actual veteran status, and 2) only affects outcomes through

the first stage channel, or through its correlation with veteran status, it is considered a valid

instrument for military service.10

Consider the following GxE interaction model:

10

The first condition is easy to verify, and standard first stage statistics (partial R2 and F-statistic) for the

significance of the instruments in the HRS sample show draft eligibility and its interaction with the GRS

are robust predictors of veteran status and its interaction with the GRS (tables are available upon request).

The exclusion restriction cannot directly be tested. In this study, a violation of the exclusion restriction

could occur if the stress of having a low draft number triggered smoking behavior. Heckman (1997)

shows the IV estimator is not consistent if heterogeneous behavioral responses to the treatment—or

military service in this case—are correlated with the instrument (i.e. draft eligibility). However, past

research has provided convincing counterfactuals that suggest the exclusion restriction holds. For

example, Angrist (1990) found no significant relationship between earnings and draft eligibility status for

men born in 1953 (where draft eligibility was defined using the 1952 lottery cutoff of 95). Since the 1953

cohort was assigned RSNs but never called to service, if the draft affected outcomes directly, we would

expect outcomes to vary by draft eligibility for this cohort.

21

(1)

Where, is coded as “1” if individual reports serving in the military and “0”

otherwise, is the genetic risk score for smoking initiation, is their interaction,

is the outcome of interest, is a vector of exogenous controls, and is the disturbance term.

The and terms are treated as endogenous and the and

terms are used as excluded instruments in a two-stage least squares (2SLS) IV

framework, where is coded as “1” for draft eligible men and “0” for draft ineligible men

according to the eligibility cutoffs reported in Table 1. For our GxE interaction model, the first-

stage regressions are as follows:

(2)

(3)

where the model is exactly identified. The vector of exogenous controls ( includes the GRS

effect for draft ineligible non-veterans, the first 20 principal components to account for

population stratification in the genotype data, and dummies for year and month of birth.11

To

increase precision in the first stage estimates, or to strengthen the correlation between the

and terms, we model the main effect of the GRS in as

, where is equal to “1” if individuals were not drafted and never served

11

A mechanical failure in the implementation of the first round of the lottery (balls with the days of the

year were not mixed sufficiently after having been dumped in a month at a time) resulted in a

disproportionately high probability of being drafted for those born in the last few months of the year

(Fienberg, 1971). This could bias estimates if those born later in the year differ in important ways from

those born at other times during the year. For example, studies have shown health varies with season of

birth.

22

in the military.12

The first stage equations are then substituted into the structural equation to

derive the reduced form in the second stage:

(4)

For all limited and continuous dependent variables in this study, the second stage

equation is estimated with a simple linear probability model because it is the ideal specification

when faced with a set of simultaneous equations where the instrument, the endogenous regressor,

and the dependent variable take on a limited set of values (Angrist & Pischke, 2008). As long as

is exogenous and only affects the outcome through the first stage channel, or through its

correlation with , we avoid any potential confounders, and the coefficient on can be

interpreted as the local average treatment effect (LATE) of military status, or the marginal effect

of veteran status at the mean GRS.13

Similarly, because genetic inheritance is by and large the

result of a random shuffling of paternal and maternal genotypes at conception, is

exogenous, and the coefficient on the GxE interaction term ( can be interpreted as the

marginal effect of veteran status by genotype. Therefore, a statistically significant coefficient on

12

In a typical linear regression model with an interaction term, the interaction term and each of the

corresponding main effects are included as separate terms (e.g. “draft”, “GRS x draft” and “GRS”). Here,

because we are using a 2SLS approach, and the “GRS” term is highly correlated with “GRS x draft”, we

model the main effect of the GRS as “nodraft x GRS” to strengthen the correlation between the “GRS x

veteran” and the “GRS x draft” terms in the first stage. Using the GxE interaction term for draft ineligible

non-veterans instead of the main effect of the GRS does not change the meaning of this term, which can

still be interpreted as the marginal effect for men who were not drafted and who did not serve. However,

it does change the interpretation of , which now represents the marginal effect for draft eligible

veterans instead of the difference between the marginal effects for draft eligible veterans and draft

ineligible non-veterans.

13

The IV estimates of effects of military service using the draft lottery capture the effect of military

service on “compliers”, or men who served because they were draft eligible but who would not otherwise

have served. It is not, therefore, an estimate of the effect of military service on men who volunteered.

See Angrist and Pischke (2008) for a more detailed discussion of the interpretation of the LATE for the

Vietnam-era draft.

23

would indicate the effect of veteran status on smoking behavior varies by genotype, or the

existence of a synergistic relationship between genotype and military service on the phenotypic

outcome of interest at higher (or lower) values of the GRS. In this way, the model allows us to

estimate the effects of military service across the entire distribution of genetic risk for smoking

initiation.

5. Results

5.1 GxE Estimates for Smoking Behavior and Smoking-Related Morbidities

Our main results are presented in Table 7, which reports OLS and 2SLS estimates from

the GxE interaction model. Results from the OLS models fail to find a significant GxE

interaction effect among veterans. However, since all omitted genetic and environmental

variables or pathways between self-reported veteran status and smoking behavior are not

accounted for in this model, we cannot rule out the possibility that the naïve estimates are biased

downward due to a host of other unobserved GxE, GxG, or ExE interactions.

Turning to the 2SLS models, we present the causal impact of veteran status on smoking

behavior and smoking-related morbidities across the entire spectrum of genetic risk for smoking.

2SLS estimates of the “Veteran" coefficient capture the effect of conscription on our outcomes

of interest at the mean GRS and the “GRS x Veteran" coefficient captures the effect of

compulsory military service at all other values of the GRS. The "GRS x Non-Veteran”

coefficient is the effect of the GRS on men who were not treated, or those who were not drafted

and who did not serve in the military. The results indicate the presence of a statistically

significant GxE interaction between Vietnam-era military service and the GRS for current or

24

former smoker status (“Ever Smoker”) and smoking quantity (“Cigarettes Per Day”). A one

standard deviation increase in the GRS of draft-induced soldiers results in a 16 percent point

increase in the likelihood of having ever been a smoker and increases quantity smoked per day

by approximately 8 cigarettes. With respect to smoking-related morbidities, we find some

evidence that veterans have a higher lifetime risk of being diagnosed with cancer and are more

susceptible to hypertension at older ages. A one standard deviation increase in the GRS

increases the risk of being diagnosed with cancer by 8.46 percent and elevates mean arterial

pressure by 4.44 units.

Given our low sample size for detecting genetic associations, we conducted power

analysis to evaluate whether the GxE coefficients ( ) reported in Table 7 are underpowered.14

Since our GRS is predictive of smoking behavior, effect sizes for are slightly larger in the

CPD (0.0161) and ever smoker (0.0088) models than they are in the more indirect, smoking-

related morbidity models for cancer (0.0054), and MAP (0.0078).15

Using these effect sizes and

the standard benchmark of 80% power , the statistical power—or the probability of

identifying a statistically significant GxE effect under the alternative hypothesis that a true

association exists—is found to be adequate for the CPD model ( . Not surprisingly,

given their lower effect sizes, the ever smoker ( ), cancer ( ) and MAP

( ) models are underpowered by conventional standards, indicating further GxE

14

Post hoc power analysis was conducted for the GxE coefficients using the software package G*Power

(Faul, Erdfelder, Buchner & Lang, 2009). The sample size of 631 was used for statistical power analysis

on a multivariate linear regression equation with 83 predictors at the conventional, two-tailed 0.05

significance level ( . 15

The partial R2 or Cohen’s effect size for the GxE coefficient was estimated from an OLS regression of

Equation 4. To minimize any potential bias in the estimation of the effect sizes due to the endogeneity of

self-reported veteran status, we model the GxE interaction terms in the OLS model using draft eligibility

status instead of self-reported veteran status.

25

analysis on a larger sample of veterans would buttress the robustness of the findings we report

here.

Across the board, we find no evidence that military service impacted the smoking

behavior or health of conscripts with an average genetic predisposition for smoking—i.e. our

findings are commensurate with other studies that are unable to identify a LATE of military

service on smoking behavior and health (Dobkin & Shabani, 2009; Eisenberg & Rowe, 2009).

These results indicate prior studies may in part have been unable to detect an overall LATE

precisely because outcomes vary considerably by genotype. To illustrate this further, the first

two rows in Table 8 estimate the total treatment effect of military service on smoking behavior

for high-risk genotypes by taking a post-estimation linear combination of the marginal effects for

veterans with GRSs one and two standard deviations above the mean.16

Veterans with a GRS

one standard deviation above the mean are 56.9 percent more likely to have ever been smokers.

Moreover, during their heaviest smoking years, these veterans smoked almost 18 more cigarettes

per day than non-veterans with the same GRS—an effect that grows as we move up the genotype

distribution. Veterans with a GRS two standard deviations above the mean were 70.5 percent

more likely to become smokers and smoked more than a pack per day—or 27 cigarettes—at their

peak compared to non-veterans with the same GRS.

The sizable variation in outcomes by genotype can be seen in Figure 1, which plots the

total difference in predicted cigarettes smoked per day for veterans versus non-veterans at all

values of the smoking initiation GRS with 95 percent confidence intervals. A strong treatment

16

Specifically, we estimate the total difference between veterans and non-veterans, or the total treatment

effect, by adding the individual treatment effects for veterans (“Veteran” + “GRS x Veteran”) and

subtracting them from the marginal treatment effect for draft ineligible non-veterans (“GRS x Non-

Veteran”). To ensure the accuracy of the standard errors, this is done using a post-estimation linear

combination.

26

effect persists for veterans with a GRS at approximately one standard deviation or higher,

whereas the total treatment effect below one standard deviation is not statistically significant.

Finally, total treatment effects are significant for the smoking behavior phenotypes only, or the

GxE effect on smoking-related morbidities is only significant at the margin.

5.2 The Influence of Post-Service Educational Attainment on Smoking

To investigate the possibility that increases in smoking behavior due to military service

were offset by post-service schooling gains from veterans’ use of the GI Bill, we also estimate

separate 2SLS models that include a control variable for education (coded as “1” if the

respondent reports obtaining a college or advanced degree, and “0” otherwise) to calculate total

effects by educational attainment (see Table 8). Education may be endogenous in the smoking

model if it is correlated with omitted variables that influence smoking behavior, such as

socioeconomic status or ability. However, if extra schooling is a downstream effect of military

service (i.e. via the GI Bill) it may be a second mechanism through which draft eligibility

influenced smoking behavior. Given the high correlation between educational attainment and

health-enhancing behaviors in the literature, post-service educational attainment may have acted

as a second environmental lever that attenuated any initial shocks to smoking behavior in young

adulthood—particularly for high-risk genotypes. 17

17

The exogenous downstream effect of educational attainment might be compromised if the high

correlation between draft eligibility and schooling is a reflection of draft avoidance behavior rather than

military service—i.e. if it is a capturing the effect of men with low draft lottery numbers who “beat the

draft” by obtaining educational deferments. Angrist and Chen (2011) find small but statistically

significant positive effects of service on educational attainment for white men born between 1948-1952.

Weighing this against the sharp decline in educational deferments during the draft lottery period, they

argue there is little evidence to support the claim that increases in schooling among draft eligible men are

due to draft avoidance behavior.

27

Consistent with our hypothesis, the statistically significant difference in smoking

behavior between high-risk veterans and non-veterans completely disappears for college-

educated men, while a strong, adverse effect persists for veterans who did not obtain a college

degree. Veterans with a GRS one standard deviation above the mean who are not college

educated were 66.8 percent more likely to become smokers and report smoking almost 22

cigarettes more per day than comparable non-veterans. Two standard deviations above the mean,

high-risk veterans without a college degree were 80.8 percent more likely to become smokers

and smoked more than a pack and a half per day (31 cigarettes) than high-risk non-veterans. The

variation in outcomes by educational attainment can be seen in Figures 2 and 3, which plot the

total effects from our CPD GxE model at all values of the GRS for veterans who did and did not

obtain a college degree. GxE effects are significant at the 0.05 level for non-college educated

men with GRSs at or above 0.4, whereas the results for college educated men are statistically

significant for men at extreme risk only (i.e. GRSs greater than or equal to 3.4). These results

indicate post-service education had a considerable moderating influence on the long-term

smoking behaviors of genetically-at-risk men who were called to service.

6. Discussion

Tobacco is the leading cause of preventable death among civilian populations and

continues to be a significant and pressing health problem within the US military. The military

spends approximately $1.6 billion a year on tobacco-related health care and lost productivity

(Wedge & Bondurant, 2009), and smoking-attributable disease accounts for 16% of all deaths

and approximately 10% of all hospital bed days (Helyer, Brehm, & Perino, 1998). Using a

polygenic score to measure genetic risk and an instrumental variables approach to operationalize

28

the environmental exposure, this study finds evidence that military-induced smoking during the

Vietnam era amplified polygenic risk for cigarette consumption. Vietnam veterans with a GRS

one to two standard deviations above the mean were approximately 57 to 71 percent more likely

to be smokers and smoked 18 to 27 more cigarettes per day more than non-veterans with the

same GRS. On the other hand, we do not find a significant treatment effect for conscripts with

GRSs at or below the mean. Heterogeneous response—by genotype—might explain why past

studies that have used the draft lotteries to investigate smoking behavior have been unable to

detect a LATE.

Limitations to these analyses should be mentioned. A significant drawback of our IV

estimation strategy, and a common criticism of the natural experiment approach to IV estimation

in general, is we cannot fully spell out the underlying theoretical relationships between military

service and smoking behavior (Angrist & Krueger, 2001). For example, we cannot unravel the

influence of specific mechanisms surrounding time in Vietnam—such as combat positions, year

of arrival, or number of tours—on cigarette consumption, making it difficult to identify what

aspects of the war experience intensified smoking behavior. In other words, there is nothing in

our IV model that explains why Vietnam-era service affected smoking behavior.

However, due to endogeneity issues, current methods being used to uncover GxE

interactions in observational studies are inadequate to support policy inference. Although we

caution that estimates from this study apply specifically to veterans from the Vietnam era, and

are therefore not externally valid or cannot be generalized to the population as a whole or even to

present-era military personnel, their high degree of internal validity may direct practitioners to

effective treatments for high-risk individuals. Thus, while inducing a military draft lottery, for

example, would not be an intervention to promote public health (nor would wholesale

29

decommissioning of military personnel be politically feasible), to the extent that the Vietnam-era

draft lottery serves as a proxy for stressful events in young adulthood, or exposure to combat,

policymakers may want to design interventions to minimize similar traumatic events. In

particular, the strong impact of college education on the attenuation of smoking behavior in

adulthood signals educational attainment may be an important pathway for mitigating other

potentially harmful GxE experiences at younger ages.

In addition, given the relatively small effect size between our GRS for smoking behavior

and our phenotypes of interest, this study would benefit from the added power of a larger

sample—particularly for the indirect smoking-related morbidities we report. For example, future

research that exploits quasi-natural experiments that that can be operationalized in the UK

Biobank or larger genotyped populations, such as variation in tobacco taxes, are needed to detect

GxE associations between smoking behavior and longer-term health problems. Yet, despite our

low sample size, we are able to detect adequately powered effects of military service on CPD or

smoking quantity, underscoring the importance of using polygenic scores or genetic associations

that benefit from prior analysis of large, well-powered consortia to estimate individual allelic

effects.

More generally, we think this study provides an example for a way forward by which

genotype can serve as the moderating prism to filter out discernable heterogeneity in treatment

effects. It is primarily in this way, we believe, that genetics and social science will be fruitfully

integrated. It is important for other scholars to not only be careful about estimating

environmental factors using econometric techniques (or randomized controlled trials) that insure

the exogeneity of the environmental regime in question but to also deploy adequate safeguards

against population stratification. That is, where possible researchers should deploy within-

30

family models to guarantee that genotype is randomly assigned or—absent pedigreed data—

should control for principle components in cross-family models as we do here.

31

Works Cited

Angrist, J. D. (1990). Lifetime earnings and the Vietnam-era draft lottery: Evidence from Social

Security Administrative Records. The American Economic Review, 313-336.

Angrist, J. D., & Chen, S. H. (2011). Schooling and the Vietnam-era GI Bill: Evidence from the

draft lottery. American Economic Journal: Applied Economics, 3(2), 96-118.

Angrist, J. D., Chen, Chen, S. H.., & Frandsen, B. R. (2010). Did Vietnam veterans get sicker in

the 1990s? The complicated effects of military service on self-reported health. Journal of

Public Economics, 94(11), 824-837.

Angrist, J. D., Chen, S. H., & Song, J. (2011). Long-term consequences of Vietnam-era

conscription: New estimates using Social Security data. The American Economic Review,

101(3), 334-338.

Angrist, J. D., & Krueger, A. B. (2001). Instrumental variables and the search for identification:

From supply and demand to natural experiments. The Journal of Economic Perspectives,

15(4), 69-85.

Angrist, J. D., & Pischke, J. S. (2008). Mostly Harmless Econometrics: An Empiricist's

Companion: Princeton University Press.

Bedard, K., & Deschênes, O. (2006). The long-term impact of military service on health:

Evidence from World War II and Korean War veterans. The American Economic Review,

176-194.

Belsky, D. W., Moffitt, T. E., Baker, T. B., Biddle, A. K., Evans, J. P., Harrington, H. L., . . .

Williams, B. (2013). Polygenic risk and the developmental progression to heavy,

persistent smoking and nicotine dependence: Evidence from a 4-decade longitudinal

study. JAMA Psychiatry, 70(5), 534-542.

Benetos, A., Safar, M., Rudnichi, A., Smulyan, H., Richard, J. L., Ducimetière, P., & Guize, L.

(1997). Pulse pressure a predictor of long-term cardiovascular mortality in a French male

population. Hypertension, 30(6), 1410-1415.

Boardman, J. D., Saint O., Jarron M., Haberstick, B. C., Timberlake, D. S., & Hewitt, J. K.

(2008). Do schools moderate the genetic determinants of smoking? Behavior Genetics,

38(3), 234-246.

Bound, J., & Turner, S. (2002). Going to war and going to college: Did World War II and the GI

Bill increase educational attainment for returning veterans? Journal of Labor Economics,

20(4), 784-815.

Card, D., & Lemieux, T. (2001). Going to college to avoid the draft: The unintended legacy of

the Vietnam War. American Economic Review, 97-102.

Centers for Disease Control & Prevention. (1994). Surveillance for selected tobacco-use

behaviors--United States, 1900-1994. MMWR. Morbidity and Mortality Weekly Report,

43(3), 1-43.

Centers for Disease Control & Prevention. (2003). Prevalence of current cigarette smoking

among adults and changes in prevalence of current and some day smoking--United

States, 1996-2001. MMWR. Morbidity and Mortality Weekly Report, 52(14), 303.

Centers for Disease Control & Prevention. (2008). Smoking-attributable mortality, years of

potential life lost, and productivity losses--United States, 2000-2004. MMWR. Morbidity

and Mortality Weekly Report, 57(45), 1226.

32

Chang C.C., Chow C.C., Tellier LCAM, Vattikuti S., Purcell S.M,. Lee J.J. (2015). Second-

generation PLINK: rising to the challenge of larger and richer datasets. GigaScience, 4.

Conley, D. (2009). The promise and challenges of incorporating genetic data into longitudinal

social science surveys and research. Biodemography and Social Biology, 55(2), 238-251.

Conley, D., & Heerwig, J. A. (2011). The war at home: Effects of Vietnam-era military service

on postwar household stability. American Economic Review, 101(3), 350-354.

Conley, D., & Heerwig, J. A. (2012). The long-term effects of military conscription on mortality:

Estimates from the Vietnam-era draft lottery. Demography, 49(3), 841-855.

Daw, J., Shanahan, M., Harris, K. M., Smolen, A., Haberstick, B., & Boardman, J. D. (2013).

Genetic sensitivity to peer behaviors 5HTTLPR, smoking, and alcohol consumption.

Journal of Health & Social Behavior, 54(1), 92-108.

De Walque, D. (2007). Does education affect smoking behaviors?: Evidence using the Vietnam

draft as an instrument for college education. Journal of Health Economics, 26(5), 877-

895.

De Walque, D. (2010). Education, information, and smoking decisions evidence from smoking

histories in the United States, 1940–2000. Journal of Human Resources, 45(3), 682-717.

Dobkin, C., & Shabani, R. (2009). The health effects of military service: Evidence from the

Vietnam draft. Economic Inquiry, 47(1), 69-80.

Eisenberg, D., & Rowe, B. (2009, February). The effect of smoking in young adulthood on

smoking later in life: Evidence based on the Vietnam era draft lottery. In Forum for

Health Economics & Policy (Vol. 12, No. 2).

Farrell, P., & Fuchs, V. R. (1982). Schooling and health: The cigarette connection. Journal of

Health Economics, 1(3), 217-230.

Faul, F., Erdfelder, E., Buchner, A., & Lang, A.G. (2009). Statistical power analyses using

G*Power 3.1: Tests for correlation and regression analyses. Behavior Research Methods,

41, 1149-1160.

Fienberg, S. E. (1971). Randomization and social affairs: The 1970 draft lottery. Science,

171(3968), 255-261.

Fletcher, J. M., & Conley, D. (2013). The challenge of causal inference in gene–environment

interaction research: Leveraging research designs from the social sciences. American

Journal of Public Health, 103(S1), S42-S45.

Fortmann, Stephen P., Todd Rogers, Karen Vranizan, William L. Haskell, Douglas S. Solomon,

& John W. Farquhar. "Indirect measures of cigarette use: expired-air carbon monoxide

versus plasma thiocyanate." Preventive Medicine 13, no. 1 (1984): 127-135.

Fuchs, V. R. (1982). Time preference and health: An exploratory study. Economic Aspects of

Health, 93.

Furberg, H., Kim, Y., Dackor, J., Boerwinkle, E., Franceschini, N., Ardissino, D., . . . Merlini, P.

A. (2010). Genome-wide meta-analyses identify multiple loci associated with smoking

behavior. Nature Genetics, 42(5), 441-447.

Glassman, A. H., Helzer, J. E., Covey, L. S., Cottler, L. B., Stetner, F., Tipp, J. E., & Johnson, J.

(1990). Smoking, smoking cessation, and major depression. JAMA, 264(12), 1546-1549.

Grimard, F., & Parent, D. (2007). Education and smoking: Were Vietnam war draft avoiders also

more likely to avoid smoking? Journal of Health Economics, 26(5), 896-926.

Grossman, M. (1972). On the concept of health capital and the demand for health. The Journal of

Political Economy, 80(2), 223-255.

33

Grossman, M. (2006). Education and nonmarket outcomes. Handbook of the Economics of

Education, 1, 577-633.

Hatziandreu, Evridiki J., John P. Pierce, Michael C. Fiore, Verner Grise, Thomas E. Novotny, &

Ronald M. Davis. "The reliability of self-reported cigarette consumption in the United

States." American Journal of Public Health 79, no. 8 (1989): 1020-1023.

Hearst, N., Newman, T. B., & Hulley, S. B. (1986). Delayed effects of the military draft on

mortality: A randomized natural experiment. The New England Journal of Medicine,

314(10), 620.

Heckman, J. (1997). Instrumental variables: A study of implicit behavioral assumptions used in

making program evaluations. Journal of Human Resources, 441-462.

Heerwig, J. A., & Conley, D. (2013). The causal effects of Vietnam-era military service on post-

war family dynamics. Social Science Research, 42(2), 299-310.

Helyer, A. J., Brehm, W. T., & Perino, L. (1998). Economic consequences of tobacco use for the

Department of Defense, 1995. Military Medicine, 163(4), 217-221.

Institute of Medicine. (1994). Veterans and agent orange: Health effects of herbicides used in

Vietnam.Washington, DC: National Academy Press.

Johnston, L. D., O'Malley, P. M., Bachman, J. G., & Schulenberg, J. E. (2005). Monitoring the

Future: National Survey Results on Drug Use, 1975-2004. Volume II: College Students

& Adults Ages 19-45, 2004. National Institutes of Health.

Kandel, D. B., & Logan, J. A. (1984). Patterns of drug use from adolescence to young adulthood:

Periods of risk for initiation, continued use, and discontinuation. American Journal of

Public Health, 74(7), 660-666.

Kendler, K. S., Gardner, C., Jacobson, K. C., Neale, M. C., & Prescott, C. A. (2005). Genetic and

environmental influences on illicit drug use and tobacco use across birth cohorts.

Psychological Medicine, 35(09), 1349-1356.

Le Marchand, L., Derby, K. S., Murphy, S. E., Hecht, S. S., Hatsukami, D., Carmella, S. G., ... &

Wang, H. (2008). Smokers with the CHRNA lung cancer–associated variants are exposed

to higher levels of nicotine equivalents and a carcinogenic tobacco-specific nitrosamine.

Cancer Research, 68(22), 9137-9140.

Liu, J. Z., Tozzi, F., Waterworth, D. M., Pillai, S. G., Muglia, P., Middleton, L., . . . Waeber, G.

(2010). Meta-analysis and imputation refines the association of 15q25 with smoking

quantity. Nature Genetics, 42(5), 436-440.

Lleras-Muney, A. (2005). The relationship between education and adult mortality in the United

States. The Review of Economic Studies, 72(1), 189-221.

Maes, H. H., Woodard, C. E., Murrelle, L., Meyer, J. M., Silberg, J. L., Hewitt, J. K., . . .

Carbonneau, R. (1999). Tobacco, alcohol and drug use in eight-to sixteen-year-old twins:

The Virginia Twin Study of Adolescent Behavioral Development. Journal of Studies on

Alcohol and Drugs, 60(3), 293.

Merline, A. C., O'Malley, P. M., Schulenberg, J. E., Bachman, J. G., & Johnston, L. D. (2004).

Substance use among adults 35 years of age: Prevalence, adulthood predictors, and

impact of adolescent substance use. American Journal of Public Health, 94(1), 96-102.

Patrick, Donald L., Allen Cheadle, Diane C. Thompson, Paula Diehr, Thomas Koepsell, and

Susan Kinne. (1994). The validity of self-reported smoking: a review and meta-analysis.

American Journal of Public Health, 84(7), 1086-1093.

Purcell, S., & Chang, C. (2014). PLINK [Computer software]. <https://www.cog-

genomics.org/plink2>.

34

Rouse, B., Sanderson, C., & Feldmann, J. (2002). Results from the 2001 National Household

Survey on Drug Abuse: Volume I. Summary of National Findings. Rockville: Substance

Abuse and Mental Health Services Administration, Office of Applied Studies.

Slomkowski, C., Rende, R., Novak, S., Lloyd‐ Richardson, E., & Niaura, R. (2005). Sibling

effects on smoking in adolescence: Evidence for social influence from a genetically

informative design. Addiction, 100(4), 430-438.

Spitz, M. R., Amos, C. I., Dong, Q., Lin, J., & Wu, X. (2008). The CHRNA5-A3 region on

chromosome 15q24-25.1 is a risk factor both for nicotine dependence and for lung

cancer. Journal of the National Cancer Institute, 100(21), 1552-1556.

Stanley, M. (2003). College education and the midcentury GI Bills. The Quarterly Journal of

Economics, 118(2), 671-708.

Thorgeirsson, T. E., Geller, F., Sulem, P., Rafnar, T., Wiste, A., Magnusson, K. P., ... &

Oskarsson, H. (2008). A variant associated with nicotine dependence, lung cancer and

peripheral arterial disease. Nature, 452(7187), 638-642.

Thorgeirsson, T. E, Gudbjartsson, D. F., Surakka, I., Vink, J. M., Amin, N., Geller, F., . . .

Walter, S. (2010). Sequence variants at CHRNB3-CHRNA6 and CYP2A6 affect

smoking behavior. Nature Genetics, 42(5), 448-453.

Timberlake, D. S., Rhee, S. H., Haberstick, B. C., Hopfer, C., Ehringer, M., Lessem, J. M, . . .

Hewitt, J. K. (2006). The moderating effects of religiosity on the genetic and

environmental determinants of smoking initiation. Nicotine & Tobacco Research, 8(1),

123-133.

U.S. Department of Health and Human Services. (1989). Reducing the health consequences of

smoking: 25 Years of Progress. A Report of the Surgeon General. U.S. Department of

Health and Human Services, Public Health Service, Centers for Disease Control, Center

for Chronic Disease Prevention and Health Promotion, Office on Smoking and Health.

DHHS Publication No. (CDC) 89-8411.

U.S. Department of Health and Human Services. (1990). A report of the surgeon general: The

health benefits of smoking cessation. Washington D.C.

U.S. Department of Veteran Affairs. (2015). Vietnam Veterans.

< http://www.benefits.va.gov/persona/veteran-vietnam.asp>.

U.S. Selective Service System. (2015). Induction Statistics. < https://www.sss.gov/induct.htm>.

Wedge, R., & Bondurant, S. (Eds.). (2009). Combating Tobacco Use in Military and Veteran

Populations. Washington D.C.: National Academies Press.

Winefield, H.R., Winefield, A. H., Tiggemann, M., & Goldney, R.D. (1989). Psychological

concomitants of tobacco and alcohol use in young Australian adults. British Journal of

Addiction, 84(9), 1067-1073.

35

Table 2. Descriptive statistics by veteran status, 1948-1952 birth cohorts, white males

Non-

veteran Veteran t Statistic

Not Draft

Eligible

Draft

Eligible t Statistic

Ever smoked GRS 0.023 -0.051 0.871 -0.130 -0.057 -0.960

(0.988) (1.028)

(0.966) (0.93)

Cigarettes per day GRS 0.178 0.467 -5.646** 0.249 0.293 -0.905

(0.623) (0.536)

(0.64) (0.576)

Ever smoker 0.577 0.692 -2.754** 0.596 0.635 -0.993

(0.495) (0.463)

(0.491) (0.482)

Current smoker 0.247 0.343 -2.517** 0.269 0.287 -0.499

(0.432) (0.476)

(0.444) (0.453)

Cigarettes per day 11.783 16.116 -2.948** 12.593 13.823 -0.891

(17.263) (16.852)

(16.919) (17.635)

Cancer 0.088 0.167 -1.461 0.115 0.110 0.360

(0.393) (0.619)

(0.478) (0.476)

Mean arterial pressure 99.843 98.246 1.491 99.723 98.879 0.845

(12.640) (11.629)

(12.729) (11.861)

Hispanic 0.171 0.045 4.385** 0.140 0.121 0.732

(0.377) (0.209)

(0.348) (0.326)

Married 0.866 0.864 0.082 0.871 0.858 0.471

(0.341) (0.344)

(0.336) (0.350)

No degree 0.143 0.040 3.854** 0.120 0.099 0.836

(0.351) (0.197)

(0.326) (0.299)

High school diploma 0.418 0.606 -4.450** 0.450 0.511 -1.520

(0.494) (0.490)

(0.498) (0.500)

Associates degree 0.035 0.121 -4.243** 0.069 0.053 0.807

(0.183) (0.327)

(0.253) (0.225)

College or adv. degree 0.404 0.232 4.252** 0.361 0.337 0.632

(0.491) (0.423)

(0.481) (0.473)

N 433 198 349 282

Notes: Estimated from the Health and Retirement Study (HRS). Standard deviations are

reported in parentheses. * p<0.05, ** p<0.01.

36

Table 3. Descriptive statistics by genotyped status, 1948-1952 birth cohorts,

white males

All

Not

genotyped Genotyped t Statistic

Ever smoker 0.633 0.657 0.613 1.538

(0.482) (0.475) (0.487)

Current smoker 0.275 0.273 0.277 -0.176

(0.447) (0.446) (0.448)

Cigarettes per day 12.032 10.733 13.143 -2.390*

(17.231) (17.145) (17.24)

Cancer 0.077 0.07 0.113 -0.771

(0.266) (0.256) (0.476)

Mean arterial pressure 99.696 101.144 99.348 1.588

(12.405) (12.577) (12.340)

Hispanic 0.155 0.183 0.132 2.443*

(0.362) (0.387) (0.338)

Married 0.846 0.824 0.865 -1.951*

(0.361) (0.381) (0.342)

No degree 0.121 0.133 0.111 1.170

(0.327) (0.34) (0.314)

High school diploma 0.481 0.485 0.477 0.279

(0.5) (0.5) (0.499)

Associates degree 0.063 0.065 0.062 0.211

(0.243) (0.246) (0.241)

College or adv. degree 0.335 0.317 0.350 -1.213

(0.472) (0.466) (0.477)

Ever smoked GRS

0

(1.00)

Cigarettes per day GRS

0

(1.00)

N 1171 540 631

Notes: Estimated from the Health and Retirement Study (HRS). Genetic risk

scores are standardized. Standard deviations are reported in parentheses.

* p<0.05, ** p<0.01.

37

Table 4. Main effect of smoking genetic risk scores on smoking behavior and smoking-

related morbidity, white males

Smoking Initiation

GRS Cigarettes

Per Day Ever smoker Cancer

Mean arterial

pressure

All white males 0.984*** 0.0341*** 0.00532 0.0321

(0.335) (0.00756) (0.00679) (0.217)

N 4297 4276 4297 4243

R2 0.0144 0.00957 0.00806 0.00718

Veteran sample 1.756** 0.0447** 0.00670 0.826

(0.722) (0.0204) (0.0117) (0.523)

N 631 631 631 620

R2 0.0506 0.0548 0.0232 0.0385

Cigarettes Per Day

GRS Cigarettes

Per Day Ever smoker Cancer

Mean arterial

pressure

All white males 1.090*** 0.0211*** -0.0102 -0.0276

(0.350) (0.00792) (0.00710) (0.227)

N 4297 4276 4297 4243

R2 0.0146 0.00649 0.00840 0.00717

Veteran sample 1.077 0.0163 0.00519 -0.258

(0.773) (0.0218) (0.0125) (0.566)

N 631 631 631 620

R2 0.0444 0.0482 0.0230 0.0349

Notes: Each column reports a separate ordinary least squares (OLS) regression of the

dependent variable on a constant and the smoking GRS. All regressions include

controls for population stratification in the genotype data and Hispanic ethnicity.

Genetic risk scores are standardized. Robust standard errors are in parentheses.

* p<0.10, ** p<0.05, *** p<0.01.

38

Table 5. Main effect of smoking initiation genetic risk score on smoking behavior and

smoking-related morbidities by draft eligibility, 1948-1952 birth cohorts, white males

Cigarettes Per

Day

Ever

smoker Cancer

Mean arterial

pressure

Draft Eligible 3.339*** 0.0688** 0.0308* 1.871**

(1.157) (0.0300) (0.0175) (0.811)

N 282 282 282 275

R2 0.133 0.170 0.123 0.127

Draft Ineligible 0.283 0.0173 -0.0106 0.304

(0.973) (0.0283) (0.0169) (0.825)

N 349 349 349 345

R2 0.0392 0.0615 0.0466 0.0374

Notes: Each column reports a separate ordinary least squares (OLS) regression of the

dependent variable on a constant and the smoking GRS. All regressions include

controls for population stratification in the genotype data and Hispanic ethnicity.

Genetic risk score is standardized. Robust standard errors are in parentheses.

* p<0.10, ** p<0.05, *** p<0.01

Table 6. OLS and 2SLS estimates of veteran status on smoking behavior and

smoking-related health outcomes, 1948-1952 birth cohorts, white males

Dependent variable OLS 2SLS

Cigarettes per day 5.230*** 8.339

(1.605) (9.621)

Ever smoker 0.147*** 0.424

(0.0459) (0.283)

Cancer 0.0424 -0.0180

(0.0264) (0.158)

Mean arterial pressure -1.666 -4.858

(1.210) (7.092)

Notes: All regressions control for respondent month and year of birth,

population stratification in the genotype data, and Hispanic ethnicity. Genetic

risk score is standardized. The sample size is 631, except for mean arterial

pressure (N=620). Robust standard errors are in parentheses.

39

Table 7. OLS and 2SLS estimates of military service and smoking genotype on smoking behavior and smoking-related

morbidities, 1948-1952 birth cohorts, white males

Cigarettes Per Day Ever Smoker Cancer Mean Arterial Pressure

OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS

Veteran 5.176*** 9.132 0.142*** 0.433 0.0439 -0.00812 -1.689 -4.531

(1.590) (9.518) (0.0450) (0.285) (0.0285) (0.154) (1.159) (7.303)

GRS x Veteran 1.982* 8.194*** 0.0220 0.161** 0.0279 0.0846* 1.201 4.438**

(1.187) (2.874) (0.0332) (0.0817) (0.0192) (0.0432) (0.950) (2.049)

GRS x Non-Veteran 1.173 -0.660 0.0540** 0.0249 -0.00602 -0.0165 0.719 0.288

(0.947) (1.075) (0.0270) (0.0351) (0.0159) (0.0175) (0.723) (0.986)

N 631 631 631 620

Notes: Estimated from the Health and Retirement Study. IV estimates of the "GRS x Veteran" coefficient capture the effect of

military service on men who served because they were draft eligible, while the "GRS x Non-Veteran” coefficient captures the

effect of the GRS on draft ineligible non-veterans. All regressions control for respondent month and year of birth, population

stratification in the genotype data, and Hispanic ethnicity. Genetic risk score is standardized. Robust standard errors are in

parentheses. * p<0.10, ** p<0.05, *** p<0.01.

40

Table 8. Total treatment effects for high-risk genotypes, 1948-1952 birth cohorts, white males

Cigarettes Per

Day Ever Smoke Cancer

Mean Arterial

Pressure

Total Veteran Effect (GRS=1) 17.99* 0.569** 0.0930 -0.380

(9.881) (0.287) (0.169) (7.679)

Total Veteran Effect (GRS=2) 26.84** 0.705** 0.194 3.770

(11.06) (0.313) (0.194) (8.605)

By educational attainment:

Total Veteran Effect (GRS=1 & College=1) 10.29 0.378 0.0779 -3.905

(12.95) (0.377) (0.222) (9.917)

Total Veteran Effect (GRS=1 & College=0) 21.96*** 0.668*** 0.101 1.482

(8.440) (0.242) (0.144) (6.657)

Total Veteran Effect (GRS=2 & College=1) 19.32 0.519 0.179 0.279

(13.79) (0.395) (0.242) (10.63)

Total Veteran Effect (GRS=2 & College=0) 30.99*** 0.808*** 0.202 5.666

(9.785) (0.273) (0.171) (7.693)

N 631 631 631 620

Notes: The total treatment effect of military service is estimated by adding the marginal treatment effects for veterans

("Veteran" + "GRS x Veteran") from the IV regressions in Table 6 and subtracting them from the marginal treatment

effect for non-veterans who were not drafted ("Non-Veteran x GRS"). To ensure the accuracy of the standard errors, this

is done using a post-estimation linear combination. The same technique is used to estimate the total effect of military

service for veterans who did and did not obtain a college or advanced degree using results from IV regressions that

include a control for college education. Robust standard errors are in parentheses. * p<0.10, ** p<0.05, *** p<0.01.

41

42

-60

-40

-20

0 20

40

60

80

-3 -2 -1 0 1 2 3 4

Pred

icte

d C

igar

ette

s P

er D

ay

Smoking Initiation Genotype Score

Figure 2. Difference in Predicted Cigarettes Per Day Veterans vs. Non-Veterans (College Educated)

43

-40

-20

0 20

40

60

80

-3 -2 -1 0 1 2 3 4

Pred

icte

d C

igar

ette

s Pe

r Day

Smoking Initiation Genotype Score

Figure 3. Difference in Predicted Cigarettes Per Day Veterans vs. Non-Veterans (Not College Educated)